AI in Supply Chain: Where It Works, Where It Stalls, and Why
In its 2026 Digital Trends in Operations survey, PwC asked 767 operations and supply chain leaders at US companies how their digital transformation was going. Eighty-five per cent said they were ahead of most competitors. Eighty-nine per cent said their technology investments had not fully delivered the expected results.
Those are the same people answering both questions.
The gap is not hypocrisy. It reflects something specific about AI in supply chain: the technology works extremely well in a few functions and stalls badly in others, and the difference is almost entirely about data that was already being collected before anyone mentioned AI.
This article covers which supply chain functions have documented results, which do not, and how to tell in advance which category yours falls into.
Why supply chain is a special case
Most enterprise functions have roughly uniform data quality. Supply chain does not. It contains both the best-instrumented processes in the business and some of the worst, often within the same department.
Demand planning has years of clean, structured, timestamped transactional history. Every order, every shipment, every stockout is in a system, with a date and a quantity, because it had to be for the business to invoice anyone.
Supplier risk does not. The signals live in contracts, emails, news, audit reports, and the knowledge of whoever manages the relationship. Nothing is structured and nothing is complete.
That difference matters more than any technology decision, because it determines whether you can find out if AI will work before committing to a build. A forecasting model can be tested against last year’s data in a week. A supplier risk model cannot be tested against anything, because the ground truth was never recorded.
This is the same asymmetry that separates predictive from generative use cases across the enterprise generally — supply chain just contains an unusually sharp example of both.
There is also external evidence that this sector is where the results land. McKinsey’s 2026 State of AI survey found that among organisations reporting cost reductions from AI at the business-function level, those reductions were most frequently reported in supply chain management, service operations and manufacturing. Enterprise-level profit impact remains rare; function-level cost impact is reported here more than almost anywhere else.
Where AI in supply chain has documented results
Demand forecasting and supply chain planning
Data readiness: high. The use case with the longest operational history and the cleanest inputs.
PwC found that AI-enabled tools are most commonly applied to planning and forecasting, at 66% of surveyed organisations, with consumer markets companies running ahead at 65% already deploying AI agents in demand planning. AI supply chain planning is the most adopted function in the category by a wide margin, and that is not fashion — it is the function where the data was already there.
What makes it work. You can backtest. Take last year’s demand, run the model, compare against what actually happened. The accuracy question is answered before the budget is committed. No other supply chain AI use case offers that.
What it still breaks on. Promotions, new product introductions, and genuine demand shocks remain hard because the historical pattern does not contain them. A model that performs well on steady-state SKUs and badly on launches is the normal outcome, not a failure — the business case should be written for the SKUs where it works.
For the underlying methods — statistical versus machine learning approaches, horizon selection, and how forecast accuracy is measured — see forecasting in supply chain management and machine learning for time series forecasting.
Inventory optimisation
Data readiness: high. Inventory positions, lead times, and service levels are all systems of record.
This is downstream of forecasting and benefits directly from it. The economics are unusually clean: better forecast accuracy reduces safety stock, which releases working capital, which is a number the CFO already reports. That makes it one of the few supply chain AI cases where the benefit appears in the ledger rather than in a slide.
What it breaks on. Lead time data quality. Many organisations record the lead time the supplier promised rather than the one actually observed, and a model trained on promises optimises for a world that does not exist.
This assumes an inventory system capable of supporting it. For the platform layer underneath — what a modern inventory stack contains and how automation is implemented — see technologies in inventory management and automated inventory systems.
Routing and logistics
Data readiness: high. Routes, times, distances and costs are all measured because they are all billed.
Fast payback relative to other supply chain AI work, because the optimisation problem is well understood, the constraint set is explicit, and the result is visible within weeks rather than quarters.
What it breaks on. Less than most functions. This is the closest thing to a safe bet in the category.
Where AI in supply chain stalls
Supplier risk and resilience
Data readiness: low. AI in supply chain risk management is the most-discussed application in the category and one of the least delivered.
The problem is not model capability. It is that the signals which would predict a supplier failure — financial distress, quality drift, management change, regional disruption — are scattered across unstructured sources, and the historical record of actual failures is too sparse to train on. Most organisations have experienced a handful of genuine supplier failures, which is not a dataset.
What does work here is narrower: monitoring and alerting against defined signals, with a human assessing. That is a useful system. It is not prediction, and the business case should not be written as though it is.
Procurement
Data readiness: mixed. PwC found 64% of organisations applying AI to sourcing and procurement, making it the second most common application — but the function spans two very different problems.
Spend classification and contract extraction are document-processing tasks with reasonable data and reasonable results — generative AI development work, with the evaluation difficulty that implies. Supplier selection and negotiation support are judgement tasks where the training signal barely exists.
Splitting those two apart before scoping is the single most useful thing a procurement AI initiative can do.
End-to-end autonomous planning (agentic AI)
Data readiness: irrelevant — the constraint is elsewhere. This is the vendor pitch increasingly marketed as “agentic AI” or “AI agents” in supply chain: a system that senses demand, adjusts supply, and reroutes shipments without human intervention.
Gartner has found that only 17% of surveyed supply chain organisations were pursuing immediate transformational redesign of processes and workflows around AI. The remainder are applying AI inside existing processes, which is the right instinct. S&P Global’s 2026 research found that across all industries only 22% of AI projects target a fully autonomous end state at all.
An initiative scoped as autonomous planning will be rescoped to decision support somewhere around month four. Scoping it that way at the start saves the rescope.
Why the published supply chain AI figures should be treated carefully
Search for supply chain AI statistics and the same numbers appear repeatedly: 87% of organisations using AI for demand forecasting, 307% average ROI within eighteen months.
Those figures circulate widely, and tracing them to a primary source is difficult. They appear on aggregator sites citing other aggregator sites, and in several cases the chain loops back on itself. That does not make them false. It does mean they should not go into a business case.
A specific example of how these figures form. A widely repeated claim holds that Unilever improved forecast accuracy from 67% to 92% through AI. Unilever has published a 92% figure — it is the reduction in frozen period at its Tinsukia factory, the time needed to replan production, cut from fourteen days to one. It has nothing to do with forecast accuracy. The number migrated between contexts and acquired a new meaning on the way. Unilever’s actual published forecasting figures from that site are more modest and more useful: demand prediction improved by 35%, finished goods inventory reduced by 16%.
The figures that do trace cleanly are less flattering and more useful:
| Finding | Source |
|---|---|
| 89% say technology investments have not fully delivered expected results | PwC, 2026 Digital Trends in Operations, 767 US operations and supply chain leaders |
| Only 30% say improvement has been significant; 58% describe it as slight | PwC, same survey |
| Data quality and access are the top barrier to ROI, ahead of legacy integration and user adoption | PwC, same survey |
| Only 37% of organisations attribute any EBIT impact to AI — 63% report none, unchanged from 2025. Just 6% attribute 5% or more | McKinsey, State of AI 2026, 1,719 respondents across 97 countries |
| Cost reductions from AI are most frequently reported in supply chain management, service operations and manufacturing | McKinsey, same survey |
| Only 22% of AI projects target a fully autonomous end state | S&P Global, AI and Labor Landscape 2026 |
The pattern across all of them: broad adoption, narrow transformation, and a measurement problem. Which is the same thing PwC’s 85-versus-89 split is describing.
The McKinsey pair is worth reading together. Enterprise-level profit impact is rare and has not improved in a year. Function-level cost reduction is reported, and supply chain is one of the three places it is reported most. Both are true, and the distance between them is the gap between a working function and a transformed business.
The data readiness test, applied to supply chain
We have watched enough of these stall to make the questions blunt on purpose — vague answers are how data gaps get discovered in month six instead of week one, and by then the budget is already spent. Before scoping any supply chain AI initiative, answer these five. The first one you cannot answer cleanly tells you where the project will stall.
- Is the outcome already measured? Forecast accuracy, inventory turns, on-time delivery, cost per shipment — a number operations already reports, with a current value.
- Does enough history exist to backtest? Two to three years of clean transactional data is the usual threshold. If the answer is no, you are building without being able to test first.
- Is the recorded value the actual value? The lead time trap: systems often hold the promised figure rather than the observed one. Check before modelling.
- Who owns each source system? Not “it’s in SAP” — a person who can approve access and answer quality questions.
- Will the output be acted on by a person or by a system? If by a system, the accuracy bar is far higher and the review design matters far more.
Questions 1 and 2 separate the functions in the first half of this article from those in the second. They are also answerable in a week, well before any budget is committed — which is the point of an AI readiness assessment rather than a discovery phase that runs alongside a build.
Find out which functions your data can actually support
Starting narrow, then widening
The pattern that works in supply chain is the same one that works everywhere else, and it is unglamorous: take one function with a measured baseline, prove it, then reuse the data and integration work on the adjacent function.
Forecasting into inventory optimisation is the natural first pair, because the second one inherits the first one’s data and the first one’s accuracy gain is what makes the second one pay. Routing is the natural standalone, because its data is clean and its payback is quick.
Supplier risk and autonomous planning are the natural fourth and fifth, not first, regardless of how prominently they feature in the vendor deck.
Where the sequence usually has to start. A frequent and uncomfortable finding is that the first project is not an AI project at all. ConnectSx, a US company working to streamline the surgical inventory chain, came to CodeIT with inventory being tracked and updated manually in spreadsheets — time-consuming, and with a high risk of error in a setting where errors are expensive.
The engagement began with a technical audit of the existing solution to establish what could realistically be built on, followed by a project plan, and then three applications: a web inventory management app, a mobile app, and a mobile UDI scanning app. The result was full visibility into inventory, including the ability to track items that had gone missing. The team grew to ten CodeIT specialists over the first year. Full write-up: ConnectSx — medical inventory management apps.
None of that is AI. All of it is the precondition. Every item in the data readiness test above — is the outcome measured, does history exist, is the recorded value the actual value — assumes a system that records those things in the first place. Where inventory lives in spreadsheets, the honest first project is the system, and the forecasting or optimisation layer comes after it has been running long enough to produce history worth modelling.
Delivery splits predictably. Forecasting, inventory and routing are predictive analytics and machine learning problems with established methods. Where neither fits an existing product and the workflow is specific enough to require it, that is custom AI development.
For organisations wanting an outside read on the sequencing, particularly where planning and procurement teams disagree on what comes first, AI advisory services cover the prioritisation without the whole programme. Where the uncertainty is narrow and specific — whether a model clears the accuracy bar on your actual data — a proof of concept answers it for a fraction of a build.
CodeIT’s wider work in the sector sits under logistics software development.
Turn a supply chain AI shortlist into a funded roadmap
FAQ
Most commonly demand forecasting and planning, which PwC found in 66% of surveyed organisations, and sourcing and procurement at 64%. Inventory optimisation and route planning follow. Supplier risk prediction and fully autonomous planning are widely discussed and far less widely delivered, because the data needed to train and test them generally does not exist in usable form.
It is the supply chain function with the strongest case, because it is the only one where you can test before you build. Years of clean transactional history let you backtest a candidate model against what actually happened, which answers the accuracy question for the cost of an analyst’s time rather than the cost of a project.
Agentic framing is mostly applied to the autonomous-planning end of the spectrum — sensing demand and rerouting shipments without human intervention — which is exactly the use case this article found least supported by evidence. Gartner found only 17% of supply chain organisations pursuing transformational redesign around AI, and S&P Global found only 22% of AI projects across all sectors targeting a fully autonomous end state. The agentic framing that has documented results is narrower: monitoring and alerting, with a person deciding.
Data quality, most often. PwC found data quality and access to be the leading barrier to ROI, ahead of legacy system integration and user adoption. McKinsey’s 2026 survey found 63% of organisations reporting no EBIT impact from AI at all, with that share unchanged from the previous year. The failures concentrate in functions where the historical record needed to validate a model was never collected.
At the function level, yes. McKinsey’s 2026 survey found that cost reductions from AI were most frequently reported in supply chain management, service operations and manufacturing. That sits alongside a much less encouraging enterprise-level finding — only 37% of organisations attribute any EBIT impact to AI — which suggests supply chain is a good place to get a function working rather than a shortcut to enterprise-wide returns.
Then the first project is the system, not the model. Forecasting and optimisation both depend on history that a spreadsheet process does not reliably produce — and the data readiness test in this article assumes something is recording outcomes in the first place. Building the system first and the AI layer second is slower to describe and considerably faster to deliver value.
Treat widely circulated figures carefully. Several commonly cited numbers — an 87% forecasting adoption rate and a 307% average ROI are the most frequent — appear mainly on aggregator sites citing one another, with no traceable primary study. Figures from PwC, McKinsey and S&P Global can be traced and are considerably more sober.
Rarely, and the data suggests most organisations have concluded the same. Gartner has found only 17% of supply chain organisations pursuing transformational process redesign around AI, and S&P Global found only 22% of AI projects across all sectors targeting a fully autonomous end state. The common and workable pattern is AI producing a recommendation and a planner deciding.
With demand forecasting, if the transactional history is clean, followed by inventory optimisation, which inherits both the data and the accuracy gain. Routing is a good standalone first project where forecasting data is not ready. Supplier risk and autonomous planning belong later, after the data foundations from the earlier work exist.